MiniMax-M2.7-JANGTQ-CRACK
on a Mac: what it actually needs
For dealign.ai. Worked out from your repo, no sign-up,
nothing to click.
8-bit
62 layers
8 KV heads
head dim 128
28 likes
463 downloads
60.7 GB
measureddownload size, summed from the
62 safetensors files in your repo
96 GB Mac
computedthe smallest machine that holds it
at 8k context, assuming MLX can wire 75% of unified memory
What it needs as context grows
4k context1.04 GB KV + 60.7 GB weights + 1.2 GB runtime = 62.9 GB, smallest Mac 96 GB
8k context2.08 GB KV + 60.7 GB weights + 1.2 GB runtime = 64.0 GB, smallest Mac 96 GB
32k context8.32 GB KV + 60.7 GB weights + 1.2 GB runtime = 70.2 GB, smallest Mac 96 GB
128k context33.29 GB KV + 60.7 GB weights + 1.2 GB runtime = 95.2 GB, smallest Mac 128 GB
KV cache = 2 tensors x 62 layers x 8 KV heads
x 128 head dim x 2 bytes (fp16) x tokens. Read straight off your
config.json. Runtime overhead is the 1.2 GB I see for Python, MLX, the
tokenizer and activations on my own runs.
Mac by Mac
| Context | Needs | 8 | 16 | 18 | 24 | 32 | 36 | 48 | 64 | 96 | 128 |
| 4k | 62.9 GB | no | no | no | no | no | no | no | no | yes | yes |
| 8k | 64.0 GB | no | no | no | no | no | no | no | no | yes | yes |
| 32k | 70.2 GB | no | no | no | no | no | no | no | no | tight | yes |
| 128k | 95.2 GB | no | no | no | no | no | no | no | no | no | tight |
Columns are unified memory in GB. "Tight" means it fits but leaves under
2 GB for anything else, so expect swap once a browser is open.
Run it
pip install mlx-lm
mlx_lm.generate --model dealignai/MiniMax-M2.7-JANGTQ-CRACK \
--prompt "hello" --max-tokens 100
What is not on this page
not measuredTokens per second. That needs a
real run on real silicon and I will not print a number I have not seen. If you want it,
reply and I will run MiniMax-M2.7-JANGTQ-CRACK on an M5 and add the result here with the
command I used.